remote sensing

Article

High-Throughput Phenotyping of Crop Water UseEfficiency via Multispectral Drone Imagery and aDaily Soil Water Balance Model

Kelly R. Thorp 1,* , Alison L. Thompson 1, Sara J. Harders 1, Andrew N. French 1 andRichard W. Ward 2

1 USDA-ARS, Arid-Land Agricultural Research Center, Maricopa, AZ 85138, USA;alison.thompson@ars.usda.gov (A.L.T.); sara.harders@ars.usda.gov (S.J.H.);andrew.french@ars.usda.gov (A.N.F.)

2 Maricopa Agricultural Center, University of Arizona, Maricopa, AZ 85138, USA; rickw@email.arizona.edu* Correspondence: kelly.thorp@ars.usda.gov; Tel.: +1-520-316-6375

Received: 12 September 2018; Accepted: 22 October 2018; Published: 25 October 2018�����������������

Abstract: Improvement of crop water use efficiency (CWUE), defined as crop yield per volume ofwater used, is an important goal for both crop management and breeding. While many technologieshave been developed for measuring crop water use in crop management studies, rarely have thesetechniques been applied at the scale of breeding plots. The objective was to develop a high-throughputmethodology for quantifying water use in a cotton breeding trial at Maricopa, AZ, USA in 2016and 2017, using evapotranspiration (ET) measurements from a co-located irrigation managementtrial to evaluate the approach. Approximately weekly overflights with an unmanned aerial systemprovided multispectral imagery from which plot-level fractional vegetation cover ( fc) was computed.The fc data were used to drive a daily ET-based soil water balance model for seasonal crop wateruse quantification. A mixed model statistical analysis demonstrated that differences in ET andCWUE could be discriminated among eight cotton varieties (p < 0.05), which were sown at twoplanting dates and managed with four irrigation levels. The results permitted breeders to identifycotton varieties with more favorable water use characteristics and higher CWUE, indicating that themethodology could become a useful tool for breeding selection.

Keywords: breeding; drought; evapotranspiration; modeling; phenomics; remote sensing; unmannedaerial system

1. Introduction

Drought stress arising from water deficit is a leading cause of crop production losses worldwide.Specific regions of the world are also projected to experience increases in the frequency and severityof future drought events due to effects of climate change on precipitation patterns [1,2]. To maintainproductivity and security of food, fiber, and other agricultural products, increasing crop water useefficiency (CWUE), defined as the ratio of crop yield and water consumption volume, is an importantgoal to ameliorate impacts of future drought. Increases in CWUE can result both from improvementsin agronomic management practices [3] and by developing improved crop cultivars that yield highlyunder water deficit conditions [4]. Although these pathways are often pursued independently withindisciplinary boundaries, the most comprehensive solutions towards CWUE improvement will beachieved through multidisciplinary efforts.

Adoption of information technologies for problem solving in the agricultural sciences has beenan especially pervasive occurrence in recent years. For example, precision irrigation, where water

Remote Sens. 2018, 10, 1682; doi:10.3390/rs10111682 www.mdpi.com/journal/remotesensing

Remote Sens. 2018, 10, 1682 2 of 19

applications are tuned based on the spatial and temporal crop need, is a primary technology forimproving CWUE in crop production settings [5,6]. Other agronomic solutions involve improvedfertility management, optimizing sowing dates, using cover crops or mulches, and choosingalternative crops or crop rotations [3]. In the plant sciences, high-throughput plant phenomics,which involves deployment of crop sensing systems for rapid phenotyping of breeding populations,has potential to guide crop improvement efforts toward cultivars with increased tolerance to droughtstress [4,7,8]. Information technologies in the form of sensors, imagers, data loggers, sensor deploymentvehicles, geographic information systems, geostatistical methods, data management approaches,and engineering control systems are central to the success of these technical solutions to improvingCWUE. However, further efforts are needed to solidify the role of information technology for practicaladvancement towards this goal.

Remote sensing has been used to estimate crop evapotranspiration (ET) and predict drought stressat multiple scales [9–13]. For example, Yebra et al. [14] combined vegetation metrics derived fromthe MODerate resolution Imaging Spectroradiometer (MODIS) satellite with the Penman-Monteithequation to estimate ET for evergreen and deciduous broadleaf forests. At the agricultural field scale,Colaizzi et al. [15] installed infrared thermometers on an overhead sprinkler irrigation machine toestimate daily ET. They proposed that the approach could improve CWUE through better irrigationscheduling. In addition, Hunsaker et al. [16] used multispectral cameras aboard a manned helicopterto map the normalized difference vegetation index (NDVI) of a cotton (Gossypium hirsutum L.) field,estimated basal crop coefficients from NDVI, and scheduled irrigation using an ET-based soil waterbalance model [17]. Results for one growing season showed that the recommendations from thescheduling model led to increased CWUE when NDVI was used to estimate basal crop coefficients.Additional field testing is needed to further realize CWUE improvements using remote sensing toguide irrigation management decisions.

Leveraging the technology previously developed for precision agriculture applications, remote andproximal sensing has recently become a highly popular tool for field-based plant phenotyping,particularly by integrating imaging equipment with unmanned aerial systems (i.e., drones) to obtain veryhigh resolution imagery of plant breeding trials. For example, Zaman-Allah et al. [18] related nitrogenstress, senescence, and yield for ten maize (Zea mays L.) hybrids to NDVI from multispectral imagescollected with a drone. In addition, Sankaran et al. [19] used a multispectral camera with a drone toestimate seedling emergence and stand density for more than 100 winter wheat (Triticum aestivum L.)varieties in Washington State. Condorelli et al. [20] measured NDVI with a drone-based multispectralcamera over 248 durum wheat (Triticum durum L.) lines at different growth stages and water regimes.A subsequent genome-wide association study (GWAS) identified 22 single quantitative trait loci (QTL)for NDVI, particularly as related to terminal drought stress. Although many important advancementshave been made, moving remote sensing data analyses beyond the NDVI will enhance the quantificationof useful phenotypes.

Few studies, if any, have aimed to directly phenotype crop water use or CWUE for purposesof breeding or genomic analysis. Rather, the focus has been to estimate related plant traits, such asspectral indices that correlate with plant water status, under both well-watered and water-limitedconditions [21–24]. A key difference is that plant water status describes the plant state at a singulartime point, whereas crop water use is a process that occurs over time. For this reason, efforts tointegrate process-based simulation models with spectral measurements of plant states may providephenotypes that better express the process-based nature of plant growth and water consumption [25].Although the integration of remote sensing data with models of radiation transfer, crop growth,and soil water balance has been widely studied for regional and field-scale mapping of crop water use,such techniques have not been applied for field-based, high-throughput plant phenomics.

The main goal of this study was to develop a methodology for phenotyping crop water useand CWUE in a cotton breeding trial, using ET estimates from a co-located irrigation managementstudy to evaluate the approach. It attempts to merge efforts from the ET remote sensing community

Remote Sens. 2018, 10, 1682 3 of 19

with efforts in the plant science community in a multidisciplinary effort toward CWUE improvement.Specific objectives were to (1) estimate fractional vegetation cover ( fc) from multispectral imagescollected over the cotton field trials with a drone, (2) use plot-level fc to derive crop water usecalculations from an ET-based daily soil water balance model, and (3) identify cotton cultivars withfavorable yield, fiber quality and water use efficiency across treatments and growing seasons.

2. Materials and Methods

2.1. Field Experiments

Cotton field experiments were conducted at the University of Arizona’s Maricopa AgriculturalCenter (MAC) near Maricopa, AZ (33.079◦N, 111.977◦W, 360 m above sea level) during the 2016and 2017 growing seasons (Figure 1). The environment in central Arizona is arid and hot,with maximum daily air temperatures regularly exceeding 38 ◦C in July and August. Growingseason precipitation from April through September amounted to 42 and 51 mm in 2016 and 2017,respectively. In comparison, short crop reference ET during the same period was 1364 and 1404 mm in2016 and 2017, respectively. Thus, cotton production required irrigation to meet evaporative demand.

2014

2015

0 25 50 75 m

Access TubesIrrigation Rate (mm)

850 - 950 950 - 1050 1050 - 1150 1150 - 1250 1250 - 1350 1350 - 1450 1450 - 1550

(a)

(b)

Figure 1. Experimental layouts for (1) a large-plot cotton irrigation management study with soilwater content measurements at 64 access tube locations (Experiment #1) and (2) a small-plot cottonstudy comparing two planting dates (P1 and P2) with eight cotton cultivars and four irrigation rates(Experiment #2) in the (a) 2016 and (b) 2017 growing seasons at Maricopa, AZ, USA. Seasonal irrigationrates (mm) for each plot area are shown with false color composite images of the field on (a) 21 July2016 and (b) 15 August 2017.

Remote Sens. 2018, 10, 1682 4 of 19

Soil variability at the 6-ha field site was characterized via a multi-year soil sampling effort at160 locations (not shown). A tractor-mounted soil sampler (model 25-TS, Giddings Machine Co.,Windsor, CO, USA) was used to collect cylindrical soil samples (0.04-m diameter × 0.4-m depth)at five incremental soil profile depths centered at 0.2, 0.6, 1.0, 1.4 and 1.8 m. Soil texture analysiswas conducted in the laboratory using the hydrometer method of Gee and Bauder [26], and the soilwater holding limits of each sample was computed from texture data using the Rosetta pedotransferfunctions [27]. For ET estimation purposes (discussed below), ordinary kriging was used to spatiallyinterpolate soil water limits at the central location of each field plot (Figure 1). Geostatistics wereconducted using the “geoR” package within the R Project for Statistical Computing software. Primarily,soil texture at the field site was sandy loam and sandy clay loam with drained upper limits between0.16 and 0.22 cm3 cm−3 and lower limits between 0.08 and 0.11 cm3 cm−3.

Data from two field experiments (Experiment #1 and Experiment #2), each with a uniqueexperimental design and purpose, were used to complete the objectives of this study. The two fieldexperiments were established on adjacent land areas (Figure 1), both receiving irrigation from the sameoverhead, lateral-move sprinkler irrigation system (Zimmatic, Lindsay Corporation, Omaha, NE, USA).Advanced technology on the irrigation machine permitted the application of variable irrigation ratesbased on georeferenced irrigation maps uploaded to the machine’s control panel. The variable-rateirrigation system permitted unique control of irrigation rates from individual drop hoses, which werespaced 1.0 m apart and located at the center of each interrow area. Nozzles were positioned to emitwater less than 1.0 m above the soil surface. For uniform soil wetting prior to cotton emergence, spraypads giving a spray diameter of approximately 5.0 m were used. After emergence, the pads werechanged to a “bubbler” style, which emitted large droplets with a 0.3-m spray diameter at the center ofeach interrow area. In addition to reducing water loss to evaporation, the bubbler pads increased thespatial accuracy of irrigation applications relative to the intended application areas delineated in thegeoreferenced irrigation maps. Spatial application error with the variable-rate irrigation machine wasestimated to be less than 2.0 m. The overhead irrigation system with variable-rate capability permittedboth field experiments to incorporate multiple irrigation rates as a primary treatment.

Experiment #1 tested cotton yield and water use responses to variable irrigation rate and timingfor one commercial cotton variety (Deltapine 1549 B2XF, Monsanto Company, St. Louis, MO, USA).Treatment plots were relatively large (12.2 m (12 cotton rows) × 30.0 m) and required a majority ofthe field area (Figure 1). The experimental design included four replications of sixteen irrigationmanagement treatments (64 plots), which involved all the possible combinations of four irrigationrates applied during two distinct periods of the growing season. The four irrigation rates were 60%,80%, 100%, and 120% of the recommendation provided by an irrigation scheduling tool based on theCSM-CROPGRO-Cotton agroecosystem model [28]. The two irrigation periods were from first squareto peak bloom (approximately the first of June through mid-July) and from peak bloom to 90% openboll (approximately mid-July through the first week of September). The various combinations of thefour rates during the two times resulted in sixteen total irrigation treatments, which led to cottongrown under a variety of soil water status conditions. The entire field area was disked and planedprior to planting. Cotton was planted with a north-south row orientation on 25 April 2016 (day of year(DOY) 116) and 18 April 2017 (DOY 108).

Of highest relevance to the present study, Experiment #1 incorporated weekly soil water contentmeasurements via a field-calibrated neutron moisture meter (model 503, Campbell Pacific Nuclear,Martinez, CA, USA). After crop emergence, steel access tubes were installed at the center of eachtreatment plot (Figure 1) using the tractor-mounted soil sampler. From mid-May to early October,the neutron moisture meter was deployed on a weekly basis (approximately 20 times per growingseason) to measure soil water content from 0.1 to 1.9 m in 0.2-m incremental depths at each access

Remote Sens. 2018, 10, 1682 5 of 19

tube. Soil water content data were used to estimate ET and deep seepage (DS) between successivemeasurement events using the soil water balance approach described by Hunsaker et al. [29]:

ET =7

∑i=1

(Di,1 − Di,n−1) +n−1

∑j=1

(Rj + Ij)− DS, (1)

DS = max (0.0,10

∑i=8

(Di,1 − Di,n−1)), (2)

where ET is the total evapotranspiration and DS is the total deep seepage occurring from thebeginning of a given measurement day (denoted as day j = 1) to the end of the day before thesuccessive measurement date (denoted as j = n − 1), Di,1 and Di,n−1 are respectively the water depthmeasurements at soil depth increment i at the beginning of day 1 and at the end of day n − 1, and Rjand Ij are respectively the precipitation and net irrigation depths received on day j. For the sakeof practicality, soil water content measurements collected on the morning of day n were used toapproximate soil water content at the end of day n − 1. Deep seepage was not permitted to be negative,meaning upflux from below the neutron access tube was considered negligible. The seasonal ETestimates from neutron moisture meter data provided important verification that the ET estimationmethodology (discussed below) provided reasonably accurate results.

Experiment #2 was a breeding trial that tested the responses of eight cotton cultivars (Table 1) totwo planting dates and four irrigation rates (Figure 1). The experimental design was a randomizedcomplete block design with planting date as the block and irrigation rate treatments nested withinthe blocks. Each cultivar was replicated three times for each irrigation treatment, totaling 96 plotsper block. The first planting date treatment was sown on 26 April 2016 (DOY 117) and 19 April 2017(DOY 109), and the second planting date treatment was sown on 18 May 2016 (DOY 139) and 10 May2017 (DOY 130). The four irrigation rates were 60%, 80%, 100%, and 120% of the recommendationprovided by a simulation tool for irrigation scheduling [28] and were administered consistently fromfirst square (early June) to 90% open boll (early September). Irrigation rates were administered uniformlywithin spatial areas containing 24 treatment plots (8 cultivars × 3 replications). Buffer areas with widthof 6.0 m (six cotton rows) and planted to a single commercial cotton variety were incorporated centrallyalong the boundaries of irrigation treatments to prevent spatial error in the variable-rate irrigationapplication from affecting cultivar treatments. In 2016, the plot dimensions for cultivar treatmentswere 3.0 m (three cotton rows) × 4.6 m with a 1.5 m buffer area between plots. In 2017, the plot sizewas reduced to 2.0 m (two cotton rows) × 4.6 m, and the buffer area between plots remained 1.5 m.Due to the relatively small size and large number of treatment plots (i.e., 192 per growing season),plot-level ET estimation via neutron moisture meter measurements and other common ET measurementmethodologies was impractical. Thus, this study aimed to estimate ET by integrating plot-specificremote sensing data and an ET-based soil water balance model with site-specific soil parameterization,providing an approach for high-throughput phenotyping of cotton water use in this breeding trial.

Table 1. Summary of eight cotton cultivars planted in the cotton breeding trial during the 2016 and2017 growing seasons at Maricopa, AZ, USA.

Abbreviation Full Name Type Source

Ark071209 Ark071209 Breeding line University of Arkansas, Keiser, AR, USAArkot9704 Arkot 9704 Released germplasm University of Arkansas, Keiser, AR, USADP1044B2RF Deltapine 1044 B2RF Commercial cultivar Monsanto Company, St. Louis, MO, USADP12R244R2 Deltapine 1441 RF Commercial cultivar Monsanto Company, St. Louis, MO, USADP1549B2XF Deltapine 1549 B2XF Commercial cultivar Monsanto Company, St. Louis, MO, USAFM958 FiberMax 958 Commercial cultivar Bayer Crop Science, Raleigh, NC, USAPD07040 Pee Dee 07040 Breeding line USDA-ARS, Florence, SC, USASiokra L23 Siokra L23 Improved variety CSIRO, Canberra, ACT, Australia

Remote Sens. 2018, 10, 1682 6 of 19

The cotton breeding trial was machine harvested with a two-row picker on 1 November 2016(DOY 306) and 8 November 2017 (DOY 312). Samples from each plot were bagged and weighedseparately in the field. Prior to the harvest date, 25 mature bolls were sampled from each plot andginned on a small research-scale cotton gin. Harvest weight, fiber turnout percentage, and plot areawere used to compute cotton fiber yield (kg ha−1) for each plot. Additionally, cotton fiber from the25 mature boll samples was sent to Cotton Incorporated (Cary, NC, USA) for analysis of fiber qualityvia High Volume Instrument (HVI) methods.

2.2. Multispectral Imaging

Remote sensing surveys from an autopiloted drone were conducted over the field site on multipleoccasions in 2016 and 2017 (Table 2). Images were collected either via a five-band multispectralcamera (RedEdge, MicaSense, Seattle, WA, USA) aboard a custom-built hexacopter or a four-bandmultispectral camera (Sequoia, Parrot, Paris, France) aboard a fixed-wing drone (eBee, SenseFly,Cheseaux-sur-Lausanne, Switzerland). The bandwidths and central wavelengths for the five-bandRedEdge camera were 20 nm at 475 nm (blue), 20 nm at 560 nm (green), 10 nm at 668 nm (red),10 nm at 717 nm (red edge), and 40 nm at 840 nm (near infrared). For the four-band Sequoia camera,bandwidths and central wavelengths were 40 nm at 550 nm (green), 40 nm at 660 nm (red), 10 nmat 735 nm (red edge), and 40 nm at 790 nm (near infrared). Both cameras featured a global shutterfor multispectral image capture and an upward facing sensor for solar irradiance characterizationduring overflight. The bit depth of images was 10 and 12 bits per channel for the Sequoia andRedEdge cameras, respectively. Flight plans were designed for 80% image overlap along flight paths.Image spatial resolution at the flight altitude of 75 m was approximately 5 cm per pixel.

Table 2. Multispectral image reconnaissance from unmanned aerial systems during the 2016 and 2017cotton growing seasons at Maricopa, AZ, USA 1.

Date DOY DAP1 DAP2 Camera

28 May 2016 149 32 10 RedEdge5 June 2016 157 40 18 RedEdge

28 June 2016 180 63 41 RedEdge5 July 2016 187 70 48 RedEdge12 July 2016 194 77 55 RedEdge21 July 2016 203 86 64 RedEdge

6 August 2016 219 102 80 RedEdge31 August 2016 244 127 105 RedEdge

8 September 2016 252 135 113 RedEdge1 October 2016 275 158 136 RedEdge

3 May 2017 123 14 −7 RedEdge24 May 2017 144 35 14 Sequoia31 May 2017 151 42 21 Sequoia6 June 2017 157 48 27 Sequoia

13 June 2017 164 55 34 Sequoia20 June 2017 171 62 41 Sequoia27 June 2017 178 69 48 Sequoia11 July 2017 192 83 62 Sequoia

1 August 2017 213 104 83 RedEdge15 August 2017 227 118 97 RedEdge22 August 2017 234 125 104 RedEdge

1 day of year, DOY; days after the first planting, DAP1; days after the second planting, DAP2.

On the morning of remote sensing surveys, four 8-m × 8-m calibration tarps (Group VIIITechnologies, Provo, UT, USA) were positioned near the cotton field. Concurrently with droneimage acquisition, ground-based radiometric measurements of each calibration tarp were collectedwith a portable field spectroradiometer (FieldSpec 3, Analytical Spectral Devices, Inc., Boulder,

Remote Sens. 2018, 10, 1682 7 of 19

CO, USA), which was sensitive to wavelengths between 350 and 2500 nm. Additional radiometricmeasurements over a calibrated, 99% Spectralon panel (Labsphere, Inc., North Sutton, NH, USA) wereused to characterize incoming solar irradiance and to compute spectral reflectance factors of each tarpduring overflight. Based on the quantum efficiency and filter functions for the multispectral cameras,the tarp spectral reflectance data were filtered to provide singular reflectance factors for each tarp,corresponding to each channel of the multispectral cameras. The filtered tarp reflectance data wereused for image calibration efforts.

A commercial photogrammetry software (Pix4DMapperPro, Pix4D S.A., Lausanne, Switzerland)was used to stitch images for each channel of the multispectral cameras. Geographic coordinates from21 ground control points around the field area were surveyed with a real-time kinematic (RTK) globalpositioning receiver and were incorporated into the image processing for georeferencing purposes.The Pix4D software was set to use camera calibration information and measurements from the upwardfacing solar irradiance sensor for initial image calibration. Additionally, the software provided anoption to input the known reflectance of one calibration target. Because reflectance measurementswere available for the calibration tarps, the darkest calibration tarp was used for the known reflectanceinput, instead of the typical reflectance panel supplied by the camera manufacturer.

Comparisons of the measured tarp reflectance data with the values from tarp pixels in theorthomosaic suggested that additional image calibration was needed following Pix4D image processing.For each multispectral image band, linear regression equations were developed to relate the measuredreflectance of all four calibration targets to the average pixel value of each target in the images.Using the band math function in a commercial remote sensing software (Environment for VisualizingImages (ENVI), Version 5.5, Exelis Visual Information Solutions, Boulder, CO, USA), the linearregression equations were applied to compute reflectance for all pixels in the image. The set of four(Sequoia) or five (RedEdge) calibrated multispectral image bands were then layer stacked, such thatfalse color composite images (e.g., Figure 1) could be inspected for accuracy of image calibrationand georeferencing.

Based on spectral information in all image channels, a supervised maximum likelihood classifierwithin the ENVI software was used to segment the composite images into bare soil and vegetationcomponents. The classifier was trained by manually selecting regions of interest associated with thetwo classes. Due to the relatively high spatial resolution from drone-based images, it was generallysimple to select multiple regions of interest that, with high certainty, represented pure spectral datafor each class. The image classification result was used to produce a binary image with values of0 for bare soil and 1 for vegetation. The binary image was loaded into a geographic informationsystem (QGIS, www.qgis.org) for calculation of zonal statistics within the boundaries of each fieldplot, as delineated with shapefile polygons (Figure 1). Based on the total vegetation pixels withineach polygon, the fractional vegetation cover ( fc) of each field plot was computed for each flight date(Table 2).

2.3. Fractional Cover Modeling

To facilitate calculations of cotton water use for each field plot (discussed below), daily fc estimateswere obtained using a Kalman smoothing algorithm [30] to integrate drone-based fc measurementswith fc from a daily logistic growth model:

d fc

dt= r fc

(1 − fc

fc,max

), (3)

where r describes the growth rate (d−1) and fc,max is the maximum value that fc attains. This differentialequation has an analytical solution:

fc =

(cer(t−t0)

1 + cfc,max

er(t−t0)

), (4)

Remote Sens. 2018, 10, 1682 8 of 19

wherec =

fc,max fc,0

fc,max − fc,0(5)

and fc,0 is fc at the initial time (t0). The model was parameterized for each field plot by adjusting r,fc,0, and fc,max to minimize error between modeled fc and the drone-based fc measurements. The t0

term was fixed as the DOY for planting each plot. Subsequently, Equation (3) with fitted parameterswas incorporated into an unscented Kalman smoothing algorithm to develop daily fc time series thatconsidered uncertainty in both the measured and modeled fc data. Kalman smoothing was especiallyimportant for characterizing fc in plots where the imposed treatments caused atypical growth patterns(Figure 2). To achieve reasonable time series as assessed visually, the fc state transition variance wasset to 0.0002 while the measurement variance was set five times higher at 0.001. A Python script wasdeveloped to conduct this analysis, which incorporated a Nelder–Mead optimizer within the “scipy”package for parameter fitting and the “pykalman” package for unscented Kalman smoothing.

100 150 200 250 300Day of Year

0.0

0.2

0.4

0.6

0.8

1.0

Frac

tiona

l veg

etat

ion

cove

r (f c

) (a)

100 150 200 250 300Day of Year

(b)

Measured Modeled Smoothed

Figure 2. Daily fractional vegetation cover ( fc) estimation using Kalman smoothing to integratedrone-based fc measurements with a daily logistic growth model for plots that received (a) therecommended irrigation schedule and (b) a schedule that underwatered early and overwatered late inthe 2017 growing season at Maricopa, AZ, USA.

2.4. Crop Water Use Estimation

The daily fc time series from Kalman smoothing were rescaled to formulate basal crop coefficient(Kcb) time series, which were used to drive an ET-based daily soil water balance model as describedin the Food and Agriculture Organization of the United Nations Paper No. 56 (FAO-56) [17]. Briefly,daily crop water use (ETc) was calculated with the following equation:

ETc = (KcbKs + Ke)ETos, (6)

where ETos is the standardized short crop reference evapotranspiration [31], Ke is the soil waterevaporation coefficient, and Ks describes the effect of water stress on transpiration. Daily ETos wascomputed using data from an Arizona Meteorological Network station (AZMET) approximately 1.2 kmfrom the field site. Required meteorological measurements included the daily minimum and maximumair temperature, dew point temperature, solar irradiance, and wind speed. Based on recommendedcomputations for maximum Kcb given environmental data for Maricopa, the Kcb time series wererescaled from fc with a minimum value of 0.15 and maximum value of 1.225. The coefficients Ke and

Remote Sens. 2018, 10, 1682 9 of 19

Ks were calculated as described in FAO-56. For Ks in particular, calculations were based on a dailysoil water balance methodology, where soil water holding limits define the total plant available water.Under non-stressed conditions with root zone soil water at the drained upper limit, Ks is 1.0 and hasno effect on ETc calculations (Equation (6)). As soil water is depleted, reductions in Ks are calculated,which lead to decreased ETc from plant transpiration. As mentioned above, the soil water holdinglimits were specified uniquely for each treatment plot based on soil sampling, soil texture analysis,pedotransfer function evaluations, and geostatistical calculations at the field site. Thus, the crop wateruse calculations were based on plot-level, site-specific information for both the soil water holdingproperties and the plant growth impacts on Kcb as characterized through drone-based fc. The soilwater status on day i was calculated as follows:

Di = Di−1 − Pi − Ii + ETci + DSi, (7)

where D is the depth of root zone soil water depletion from the drained upper limit at the end of theday, P and I are the precipitation and irrigation received on the day, ETc is the daily crop water usefrom Equation (6), and DS is the water lost to deep seepage, which is calculated when water inputsexceed the root zone soil water storage capacity. This ET-based soil water balance model, as describedin FAO-56 [17], was coded as a Python script for use in all cotton water use calculations in this study.

For each treatment plot in both field experiments in both growing seasons, ETc calculations wereinitiated on the day of planting and terminated after crop maturity in the first week of October.To summarize the seasonal crop water use, daily ETc calculations were aggregated over time.For comparison to measured ETc data in the irrigation management study (Equation (1)), daily ETc

calculations were aggregated from the first to the last neutron moisture meter measurement dates ineach growing season. For the cotton breeding trial, ETc data were aggregated over the entire calculationperiod from planting to crop maturity. Furthermore, seasonal CWUE (kg m−3) was calculated fromthe ratio of measured cotton fiber yield and modeled seasonal cotton ET. To eliminate the effect ofsurface soil water evaporation, seasonal water use efficiency was also computed considering only thetranspiration component of ET (TWUE, kg m−3).

2.5. Statistical Analysis

Statistical software (SAS for Windows, ver. 9.3, SAS Institute, Cary, NC, USA) was used to fit alinear model for each of eleven cotton traits in response to the experimental treatments in the breedingtrial. The eleven cotton traits were fiber yield (FY, kg ha−1), seasonal evapotranspiration (ET, mm),crop water use efficiency (CWUE = FY:ET, kg m−3), seasonal transpiration (T, mm), transpiration wateruse efficiency (TWUE = FY:T, kg m−3), micronaire (MIC, unitless), upper half mean fiber length (UHM,mm), uniformity index (UI, mm mm−1), fiber strength (STR, HVI g tex−1), fiber elongation at failure(ELO, %), and short fiber content (SFC, %). Mixed model analysis of variance was conducted for eachtrait using the SAS MIXED procedure with the restricted maximum likelihood (REML) method. In thecase that more than one trait estimate was available per plot, the plot mean was used as the dependentvariable. The statistical model for each trait (Y) was

Yijklmn = µ + αi + β j + γk + δl(k) + (αβ)ij + (αγ)ik + (βγ)jk + (αδ)il(k) + (βδ)jl(k)

+(αβγ)ijk + (αβδ)ijl(k) + ζm + (αζ)im + εn(ijklm),(8)

where experimental year (αi), cotton cultivar (β j), planting date (γk), irrigation treatment nested withinplanting date (δl(k)), and corresponding interaction terms were fitted as fixed effects, and the replicate(ζm) and year by replicate interaction ((αζ)im) were fitted as random effects. To examine associationsamong traits, Pearson’s correlation coefficients (r) were determined using the SAS CORR procedure.

Remote Sens. 2018, 10, 1682 10 of 19

3. Results

3.1. Crop Water Use Evaluation

Evaluation of seasonal cotton water use estimates from the drone-based ET modeling approach ascompared to ET measurements from neutron moisture meters demonstrated reasonable performancewith the modeling technique (Figure 3). The root mean squared errors (RMSE) between measured andmodeled seasonal ET among 64 treatment plots in the irrigation management study were less than 5%.In both seasons, the model underestimated ET for the four treatment plots with the highest ET amounts.These treatment plots received 120% of the recommended irrigation rate over the whole season andwere therefore overwatered. One possibility is that measured ET from the neutron moisture meterswas overestimated, which is likely considering that deep seepage was estimated from soil moisturestates typically one week apart (Equation (2)). In actuality, more water flux to deep seepage may haveoccurred for the overwatered plots, which would reduce estimates for measured ET (Equation (1))if it could be accurately measured. For other treatment plots, which were not overwatered for theentire season, the agreement between measured and modeled ET was very good. In 2016 and 2017,measured deep seepage ranged from 45 to 97 mm and from 21 to 63 mm, respectively (not shown).Modeled deep seepage in both years ranged from 0 to 146 mm with a majority of plots modeled ashaving zero seepage. On average, measured deep seepage was less than 7% and 4% of ET in 2016 and2017, respectively, which demonstrates the high importance of crop water use as a pathway for waterloss from the agroecosystem. Overall, the results show that the drone-based fc and soil water balancemodeling approach could effectively estimate seasonal cotton water use at this field site.

0.6 0.7 0.8 0.9 1.0 1.1 1.2Measured ET (m)

0.6

0.7

0.8

0.9

1.0

1.1

1.2

Mod

eled

ET

(m)

(a)

2016: RMSE=4.5%

0.6 0.7 0.8 0.9 1.0 1.1 1.2Measured ET (m)

(b)

2017: RMSE=4.0%

Figure 3. Modeled versus measured cotton evapotranspiration (ET) for the 64 irrigation managementplots from day of year (a) 138 to 277 in 2016 and (b) 129 to 274 in 2017 at Maricopa, AZ, USA.

3.2. Basal Crop Coefficients

The basal crop coefficient (Kcb) time series for each plot in the breeding trial were variableaccording to the combined effect of planting date, irrigation rate, and cotton cultivar on fc, as measuredusing drone-based images (Figure 4). For example on DOY 184 in 2016, which is approximately thedate of first flower, the Kcb among plots in the first planting varied from 0.40 to 1.09 (Figure 4a).This means that some plots were transpiring as low as 23% of full transpiration capacity on this day,while others were transpiring as high as 88% of full capacity. As compared to the standard trapezoidalKcb time series, which were computed based on days after planting as recommended in FAO-56,many plots reached full transpiration capacity earlier than the standard, while others never reached

Remote Sens. 2018, 10, 1682 11 of 19

full transpiration capacity. The drone-based fc data also suggested that late season Kcb remainedhigher than the standard Kcb recommendation, which may be true because cotton was traditionally aperennial plant and for agronomic purposes requires defoliation by desiccant application to terminateleaf growth and transpiration. Overall, to the extent that Kcb via drone-based fc correlates to planttranspiration rates, the data suggest highly variable water use patterns among the treatments imposedin the cotton breeding trial.

0.2

0.4

0.6

0.8

1.0

1.2

K cb

(a) (b)

100 136 172 208 244 280Day of Year

0.2

0.4

0.6

0.8

1.0

1.2

K cb

(c)

100 136 172 208 244 280Day of Year

(d)

Curves from drone-based fc Standard curve

Figure 4. Basal crop coefficients (Kcb) as estimated from drone-based fractional vegetation cover ( fc)in the (a) first and (b) second planting in 2016 and the (c) first and (d) second planting in 2017 forbreeding trials conducted in Maricopa, AZ, USA. The standard trapezoidal Kcb curve as recommendedin FAO-56 is presented for comparison.

3.3. Statistical Results

Mean cotton fiber yield was 1831 kg ha−1 and ranged from 703 to 2929 kg ha−1 (Table 3),as measured in the cotton breeding trial for two growing seasons, two planting dates, and four irrigationrates. Computations of seasonal ET with the drone-based fc and soil water balance modeling approach

Remote Sens. 2018, 10, 1682 12 of 19

ranged from 661 to 1223 mm, indicating that some plots used up to twice as much water as others.Estimates of the seasonal transpiration component of ET ranged from 389 to 1069 mm. On average,0.19 kg of cotton fiber was produced per m3 of water consumed (i.e., CWUE). Considering onlythe transpiration component of ET, mean TWUE was 0.25 kg m−3, a 26% increase as compared tocalculations that included surface evaporation. This establishes an upper limit on the potential forCWUE improvements through management practices that reduce surface evaporation.

The fiber quality measurements suggested desirable fiber characteristics in some cases and lessdesirable characteristics in others. With micronaire values above 4.3, none of the plots produced fiberwith micronaire in the range required for premium pricing (i.e., 3.7 to 4.2). Upper half mean lengthranged from 26 to 32 mm, indicating medium to long fiber lengths. Uniformity indices ranged from 79to 86 mm mm−1, indicating average to high fiber length uniformity. Fiber strength ranged from 27 to40 HVI g tex−1, indicating average to very strong fibers. Elongation at failure ranged from 4.0% to 7.6%,indicating very low to high fiber elongation and an overall wide range for this study. With r equal to0.31 and 0.33 (not shown), CWUE and TWUE, respectively, were correlated with fiber elongation atfailure. Thus, the plots with higher water use efficiency tended to produce fibers that could stretchmore before breaking. In addition, with r equal to 0.30 and 0.34, seasonal ET and T, respectively, werecorrelated with upper half mean fiber length. Thus, plots with higher water use generally producedlonger fibers. Fiber yield did not correlate with any fiber quality metrics with r greater than 0.24. Thus,CWUE and TWUE were better estimators of fiber quality metrics than fiber yield in some cases.

Table 3. Simple statistics for eleven cotton traits as measured for the 2016 and 2017 cotton breedingtrials at Maricopa, AZ, USA, including cotton fiber yield (FY), seasonal evapotranspiration (ET),crop water use efficiency (CWUE = FY:ET), seasonal transpiration (T), transpiration water use efficiency(TWUE = FY:T), micronaire (MIC), upper half mean fiber length (UHM), uniformity index (UI),fiber strength (STR), fiber elongation at failure (ELO), and short fiber content (SFC).

Trait Unit n Mean Std Dev Minimum Maximum

FY kg ha−1 383 1831.00 445.16 703.15 2929.00ET mm 384 954.73 145.07 661.40 1223.00CWUE kg m−3 383 0.19 0.04 0.10 0.31T mm 384 757.50 158.85 389.20 1069.00TWUE kg m−3 383 0.25 0.05 0.13 0.40MIC unitless 384 5.17 0.30 4.30 6.00UHM mm 384 28.96 1.02 26.16 31.50UI mm mm−1 384 82.79 1.06 79.30 85.80STR HVI g tex−1 384 31.85 1.86 27.20 39.70ELO % 384 5.87 0.82 4.00 7.60SFC % 384 8.23 0.60 6.70 10.70

Linear mixed modeling identified many differences (p < 0.05) for the eleven cotton traits inresponse to the treatments imposed with the cotton breeding trial (Table 4). Mean FY, ET, CWUE,T and TWUE were different in the two growing seasons. However, among the fiber quality traits,only UI was different with growing season. The two planting date treatments led to differencesin most of the cotton traits, with the exception of two fiber quality traits, STR and SFC. Likewise,irrigation rate led to differences in all traits except for the STR fiber quality trait. Differences amongthe eight cotton cultivars were found for all eleven cotton traits. Fiber yield and ET were different forall the interaction terms included in the model (Equation (8)), while the remaining cotton traits wereoccasionally different for these interactions (not shown). Most importantly, the mixed modeling resultsshow that the methodology used to estimate cotton water use and CWUE led to differences for thesetraits among the eight cultivars. Thus, the approach may have value for breeding programs that aimto improve CWUE and identify drought tolerant cultivars.

Remote Sens. 2018, 10, 1682 13 of 19

Table 4. Linear mixed modeling results for the cotton breeding trial at Maricopa, AZ, USA, including Fstatistics and probability values (p) for cotton fiber yield (FY), seasonal evapotranspiration (ET),crop water use efficiency (CWUE = FY:ET), seasonal transpiration (T), transpiration water use efficiency(TWUE = FY:T), micronaire (MIC), upper half mean fiber length (UHM), uniformity index (UI),fiber strength (STR), fiber elongation at failure (ELO), and short fiber content (SFC).

Trait Year Planting Date Irrigation Rate Cultivar

F p F p F p F p

FY 21.87 0.0001 273.97 0.0001 42.98 0.0001 5.31 0.0001ET 96.23 0.0001 5589.25 0.0001 3073.79 0.0001 2.25 0.0305CWUE 35.22 0.0001 38.79 0.0001 13.66 0.0001 5.22 0.0001T 127.66 0.0001 1272.99 0.0001 941.46 0.0001 2.91 0.0061TWUE 39.42 0.0001 23.06 0.0001 41.59 0.0001 5.77 0.0001MIC 2.71 0.3474 11.35 0.0009 4.06 0.0007 15.70 0.0001UHM 10.83 0.0525 18.56 0.0001 28.66 0.0001 6.75 0.0001UI 54.69 0.0001 6.84 0.0112 11.31 0.0001 2.81 0.0079STR 0.01 0.9328 0.17 0.6776 5.19 0.0001 14.24 0.0001ELO 0.53 0.5479 8.75 0.0034 1.86 0.0884 17.84 0.0001SFC 0.24 0.6363 3.63 0.0580 12.09 0.0001 3.82 0.0006

3.4. Relevance to Breeding

Improving CWUE via plant breeding requires the identification of cultivars that yield highlybut with reduced transpiration. An important aspect of the modeling methodology is that the cropcoefficients (Equation (6)) permit partitioning of water use among plant transpiration and surfaceevaporation components. Thus, for purposes of breeding, the results can be discussed in terms ofTWUE, the crop yield per volume of water transpired because breeding will primarily improve CWUEthrough the transpiration component of ET.

Plots of fiber yield versus seasonal transpiration demonstrated a positive correlation for allgrowing season and planting date combinations (Figure 5), generally equating higher yield with higherwater use as commonly shown in crop water production functions [3]. Coefficients of determination (r2)between fiber yield and seasonal transpiration were higher with the second planting than with the firstplanting, indicating water management was a stronger driver of yield variability with later planting.Both fiber yield and seasonal transpiration were reduced with the second planting as compared to thefirst planting. The effect was due in part to a shorter growing season for the second planting because itwas planted approximately one month later than the first planting while irrigation for both plantingdate treatments was terminated at the same time in early September. Extending the irrigation periodlater into September may have increased fiber yield and transpiration for the second planting.

Reduction in seasonal transpiration for the second planting may also be due to environmentaldifferences. In central Arizona, the month of June (DOY 152 to 181) typically has high ambienttemperatures and low humidity, which leads to increased evaporative demand (i.e., ETos). In addition,as seen in Figure 4, the Kcb (i.e., transpiration rate) during this period was much higher for the firstplanting than for the second planting due to higher fc. Thus, water use in the month of June wasmuch greater for the first planting (Equation (6)). With the second planting, the hot and dry period inJune was largely avoided, and maximum Kcb was not attained until July when the Arizona monsoonseason brought higher humidity, reduced ambient air temperatures, and lower evaporative demand.Thus, improvements in CWUE may come from better timing of the maximum Kcb period with theArizona monsoon season. Furthermore, because r2 between fiber yield and transpiration was greaterfor the second planting, the Arizona monsoon season may provide a better environment for selectingheritable traits that improve yield and CWUE and increase genetic gain in breeding populations.

Remote Sens. 2018, 10, 1682 14 of 19

0.0

0.5

1.0

1.5

2.0

2.5

3.0Fi

ber Y

ield

(Mg

ha1 )

(a)

r2 = 0.349p < 0.0001

(b)

r2 = 0.546p < 0.0001

0.3 0.5 0.7 0.9 1.1Seasonal Transpiration (m)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

Fibe

r Yie

ld (M

g ha

1 )

(c)

r2 = 0.070p = 0.0094

0.3 0.5 0.7 0.9 1.1Seasonal Transpiration (m)

(d)

r2 = 0.438p < 0.0001

Figure 5. Cotton fiber yield versus seasonal transpiration for 96 treatment plots (4 irrigation rates ×8 cotton cultivars × 3 replications = 96 plots) with (a) the first planting date in 2016; (b) the secondplanting date in 2016; (c) the first planting date in 2017; and (d) the second planting date in 2017.Data were collected as part of a cotton breeding trial conducted at Maricopa, AZ, USA.

The calculated TWUE for each cotton variety in the breeding trial showed significant variationamong planting dates and irrigation rates (Figure 6). The TWUE for the first 2017 planting at the60% irrigation rate was unreasonably high (Figure 6c). This was likely due to errant irrigationapplications to those plots, which was supported by soil moisture measurements (data not shown).After removing this treatment from consideration, DP1044B2RF (C3) and Arkot9704 (C2) demonstratedthe most consistent TWUE across years, planting dates, and irrigation treatments. Furthermore, C3 wasranked as the second highest yielding line and also had consistent fiber quality traits across yearsand treatments. This means producers should expect consistent yields with C3 and C2 over a rangeof soil water status conditions. However, selecting on yield stability alone can be unreliable due tothe effects of environmental variation across growing seasons. Calculated TWUE may provide analternative phenotype for selection because aggregated environmental conditions are included in theestimate, thus reducing selection error caused by environment or management variation. For example,DP1044B2RF (C3) and FM958 (C5) consistently achieved higher TWUE at the 80% irrigation level ascompared to the higher irrigation levels, which may highlight commercial breeders’ efforts to select fordrought tolerance in the development of these varieties.

Remote Sens. 2018, 10, 1682 15 of 19

0.1

0.15

0.2

0.25

0.3

0.35

0.4TW

UE (k

g m

3 )(a) (b)

C1 C2 C3 C4 C5 C6 C7 C8Cultivar

0.1

0.15

0.2

0.25

0.3

0.35

0.4

TWUE

(kg

m3 )

(c)

C1 C2 C3 C4 C5 C6 C7 C8Cultivar

(d)

60% Rate 80% Rate 100% Rate 120% Rate

Figure 6. Transpiration water use efficiency (TWUE) for eight cotton cultivars (C1, Ark071209; C2,Arkot9704; C3, DP1044B2RF; C4, DP12R244R2; C5, DP1549B2XF; C6, FM958; C7, PD07040; C8,Siokra L23) at four irrigation rates (60%, 80%, 100%, and 120% of recommended amounts) for (a) thefirst planting date in 2016; (b) the second planting date in 2016; (c) the first planting date in 2017;and (d) the second planting date in 2017. Data were collected as part of a cotton breeding trial conductedat Maricopa, AZ, USA.

4. Discussion

A main advantage of the ET-based soil water balance model (Equations (6) and (7)) is itspartitioning of crop water use into surface evaporation and plant transpiration components viacrop coefficients (Kcb, Ks, and Ke). As dictated by the model equations [17], Ks and Ke are mainlyinfluenced by soil water holding characteristics, and Ke is additionally influenced by the irrigationsystem and ground cover. Thus, reductions in crop water use via changes to these coefficients canbe achieved only through adjustments in management practices. Some options include switching tosubsurface drip irrigation to limit surface evaporation, adopting conservation tillage to increase groundcover, or using variable-rate irrigation to adjust irrigation rates based on the site-specific soil waterholding characteristics. The remaining coefficient, Kcb, is the primary driver of plant transpiration inthe model. As such, reductions in crop water use via changes to Kcb can be achieved mainly throughplant breeding. The goal for breeding should be to select varieties that yield highly while exhibitingKcb time series (Figure 4) that reduce water use through transpiration. By incorporating models of

Remote Sens. 2018, 10, 1682 16 of 19

cropping system processes into crop breeding and management efforts, the improvements offered bymanagement versus breeding can be better quantified.

The present study achieved high-throughput phenotyping of crop water use efficiency usingrelatively simple imaging and modeling technologies. Obviously, plant transpiration is dictated byprocesses with higher complexity than can be described through fc time series alone. For example,the processes of root water uptake, water transport through the plant, leaf stomatal conductance,and other plant water regulatory mechanisms are not considered in this study. Incorporation of morecomplex models that consider plant water status in greater detail may improve transpiration estimates,although models with higher complexity often also increase the model parameterization requirements.To be effective, the model parameters would need to be accurately estimated, or perhaps phenotyped,for each variety, which could be the subject of future studies. In this study, a simpler model waschosen due to the relative ease of driving the model with basic fc measurements from drone-basedmultispectral imaging. A main contribution is the use of time-series fc estimates with a simple soilwater balance model to provide seasonal transpiration estimates for breeding trials at the plot-level.As such, the relatively simple methodology provides a practical way to incorporate information oncrop water use into decisions on breeding selection and may be especially useful when the goal is tosimply determine which varieties to keep in the breeding program.

As this study represents one of several initial efforts with drone-based imaging at the Maricoparesearch station, much has been learned over the duration of the study. Regarding the choice ofsensors, the commercial multispectral cameras (i.e., Micasense RedEdge and Parrot Sequoia) werepopular products marketed for the agricultural sector at the time of the study, and they were thereforegiven primary focus for flight hours. As such, the data from these sensors more fully covered thegrowing season as required for time-series fc estimates herein. An astute reader will note that fc timeseries could perhaps be more easily attained from digital color (RGB) cameras, which is a simplerand more pervasive imaging technology with greater commercial options for choices of camera anddrone platform. Additionally, the image spatial resolution is often higher for RGB cameras, and imagecalibration methods are simpler as compared to multispectral image calibration. Data from RGBcameras were also collected over the field site but with less frequency and consistency, which madethese data less useful for the study. Because the Parrot Sequoia camera collected both RGB andmultispectral data, efforts with this camera in 2017 (Table 2) aimed to collect both types of data ina single drone flight. However, the success of this approach was reduced by data reliability issueswith the camera and multiple crashes of the fixed-wing drone for which the camera was designed.Hence, the MicaSense RedEdge camera on a custom hexacopter drone was the favored configurationherein. For future efforts to map fc time series, an even better configuration may be an RGB cameraon a copter-style drone, for which there are now many commercial options. However, future studiesshould also further contrast the advantages and disadvantages of multispectral versus RGB imagingfrom drone platforms.

Missing from this study is the use of thermal cameras to quantify differences in plant water statusamong varieties, which again was related to giving primary focus for flight hours to the multispectralcameras. Incorporating canopy temperature information from thermal cameras into the methodologyhas great potential to improve transpiration estimates among varieties. For example, Kullberg et al. [32]recently described the use of stationary infrared thermometry to adjust Ks in the ET-based soil waterbalance model (Equation (6)), applied to maize. Although their data was collected from ground-basedsystems rather than drones, their approach demonstrates the integration of Kcb from RGB data andKs from thermal data into the ET-based soil water balance model for quantifying maize water use.Future efforts should evaluate this data fusion technique for improved high-throughput phenotypingof crop water use among crop varieties, likely using RGB and thermal data from drone-based imagesor sensors mounted on ground-based phenotyping vehicles [25].

Remote Sens. 2018, 10, 1682 17 of 19

5. Conclusions

High-throughput phenotyping of water use in 192 cotton breeding plots was achieved usingapproximately weekly vegetation cover estimates from drone-based multispectral cameras to drivebasal crop coefficients in a daily soil water balance model. Differences in water use estimates amongeight cotton cultivars were found via linear mixed modeling, indicating that the approach can assistbreeders in the identification of cultivars with favorable water use characteristics. Combining thewater use estimates with crop yield can quantify crop water use efficiency, allowing breeders toselect varieties that maximize crop production per volume of water used. Options for improvingthe methodology include considering models with greater complexity in crop water use processes,assessing alternative options for drone-based image data collection, and incorporating other cropphenotypes such as canopy temperature. Taken together, the study demonstrates a methodology forhigh-throughput phenotyping of crop water use, providing breeders a tool for water use quantificationand additional traits for consideration in selective breeding.

Author Contributions: Conceptualization, K.R.T. and A.L.T.; Methodology, All authors; Software, K.R.T., S.J.H.,and A.N.F.; Validation, All authors; Formal Analysis, K.R.T. and A.L.T.; Investigation, All authors; Resources,R.W.W.; Data Curation, K.R.T.; Writing—Original Draft Preparation, K.R.T. and A.L.T.; Writing—Reviewand Editing, All authors; Visualization, K.R.T. and A.L.T.; Supervision, K.R.T., A.N.F., and R.W.W.; ProjectAdministration, K.R.T. and R.W.W.; Funding Acquisition, K.R.T. and R.W.W.

Funding: This research was partially funded by Cotton Incorporated.

Acknowledgments: The authors acknowledge Mark Yori from Phoenix Drone Services for collection of multispectraldrone imagery for this research. In addition, the authors acknowledge the USDA-ARS-ALARC technicians andsummer employees for conducting field experiments and collecting field data to support this research, includingSuzette Maneely, Bill Luckett, Matt Hagler, Melissa Stefanek, Cassie Farwell, Josh Cederstrom, Matthew McGuire,Dusti Baker, and Joseph Griffin.

Conflicts of Interest: The authors declare no conflict of interest. The founding sponsors had no role in the designof the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in thedecision to publish the results.

References

1. Dai, A. Drought under global warming: A review. Wiley Interdiscip. Rev. Clim. Chang. 2011, 2, 45–65.[CrossRef]

2. Wehner, M.; Easterling, D.R.; Lawrimore, J.H.; Heim, R.R.; Vose, R.S.; Santer, B.D. Projections of futuredrought in the continental United States and Mexico. J. Hydrometeorol. 2011, 12, 1359–1377. [CrossRef]

3. Ali, M.H.; Talukder, M.S.U. Increasing water productivity in crop production—A synthesis. Agric. Water Manag.2008, 95, 1201–1213. [CrossRef]

4. Passioura, J.B. Phenotyping for drought tolerance in grain crops: When is it useful to breeders?Funct. Plant Biol. 2012, 39, 851–859. [CrossRef]

5. Evans, R.G.; King, B.A. Site-specific sprinkler irrigation in a water-limited future. Trans. ASABE 2012,55, 493–504. [CrossRef]

6. Sadler, E.J.; Evans, R.G.; Stone, K.C.; Camp, C.R. Opportunities for conservation with precision irrigation.J. Soil Water Conserv. 2005, 60, 371–379.

7. Araus, J.L.; Cairns, J.E. Field high-throughput phenotyping: The new crop breeding frontier. Trends Plant Sci.2014, 19, 52–61. [CrossRef] [PubMed]

8. Chapman, S.C.; Chakraborty, S.; Dreccer, M.F.; Howden, S.M. Plant adaptation to climatechange–opportunities and priorities in breeding. Crop Pasture Sci. 2012, 63, 251–268. [CrossRef]

9. Gago, J.; Douthe, C.; Coopman, R.E.; Gallego, P.P.; Ribas Carbo, M.; Flexas, J.; Escalona, J.; Medrano, H.UAVs challenge to assess water stress for sustainable agriculture. Agric. Water Manag. 2015, 153, 9–19.[CrossRef]

10. Gowda, P.H.; Chavez, J.L.; Colaizzi, P.D.; Evett, S.R.; Howell, T.A.; Tolk, J.A. ET mapping for agriculturalwater management: Present status and challenges. Irrig. Sci. 2008, 26, 223–237. [CrossRef]

11. Khanal, S.; Fulton, J.; Shearer, S. An overview of current and potential applications of thermal remote sensingin precision agriculture. Comput. Electron. Agric. 2017, 139, 22–32. [CrossRef]

Remote Sens. 2018, 10, 1682 18 of 19

12. Maes, W.H.; Steppe, K. Estimating evapotranspiration and drought stress with ground-based thermal remotesensing in agriculture: A review. J. Exp. Biol. 2012, 63, 4671–4712. [CrossRef] [PubMed]

13. Wesseling, J.G.; Feddes, R.A. Assessing crop water productivity from field to regional scale. Agric. WaterManag. 2006, 86, 30–39. [CrossRef]

14. Yebra, M.; Van Dijk, A.; Leuning, R.; Huete, A.; Guerschman, J. Evaluation of optical remote sensing toestimate actual evapotranspiration and canopy conductance. Remote Sens. Environ. 2013, 129, 250–261.[CrossRef]

15. Colaizzi, P.D.; O’Shaughnessy, S.A.; Evett, S.R.; Mounce, R.B. Crop evapotranspiration calculation usinginfrared thermometers aboard center pivots. Agric. Water Manag. 2017, 187, 173–189. [CrossRef]

16. Hunsaker, D.J.; French, A.N.; Waller, P.M.; Bautista, E.; Thorp, K.R.; Bronson, K.F.; Andrade Sanchez, P.Comparison of traditional and ET-based irrigation scheduling of surface-irrigated cotton in the aridsouthwestern USA. Agric. Water Manag. 2015, 159, 209–224. [CrossRef]

17. Allen, R.G.; Pereira, L.S.; Raes, D.; Smith, M. FAO Irrigation and Drainage Paper No. 56. Crop Evapotranspiration:Guidelines for Computing Crop Water Requirements; Food and Agriculture Organization of the United Nations:Rome, Italy, 1998.

18. Zaman-Allah, M.; Vergara, O.; Araus, J.L.; Tarekegne, A.; Magorokosho, C.; Zarco-Tejada, P.J.; Hornero, A.;Albà, A.H.; Das, B.; Craufurd, P.; et al. Unmanned aerial platform-based multi-spectral imaging for fieldphenotyping of maize. Plant Methods 2015, 11, 35. [CrossRef] [PubMed]

19. Sankaran, S.; Khot, L.R.; Carter, A.H. Field-based crop phenotyping: Multispectral aerial imaging forevaluation of winter wheat emergence and spring stand. Comput. Electron. Agric. 2015, 118, 372–379.[CrossRef]

20. Condorelli, G.E.; Maccaferri, M.; Newcomb, M.; Andrade-Sanchez, P.; White, J.W.; French, A.N.; Sciara, G.;Ward, R.; Tuberosa, R. Comparative aerial and ground based high throughput phenotyping for the geneticdissection of NDVI as a proxy for drought adaptive traits in durum wheat. Front. Plant Sci. 2018, 9, 893.[CrossRef] [PubMed]

21. El-Hendawy, S.E.; Hassan, W.M.; Al-Suhaibani, N.A.; Schmidhalter, U. Spectral assessment of droughttolerance indices and grain yield in advanced spring wheat lines grown under full and limited waterirrigation. Agric. Water Manag. 2017, 182, 1–12. [CrossRef]

22. Rischbeck, P.; Elsayed, S.; Mistele, B.; Barmeier, G.; Heil, K.; Schmidhalter, U. Data fusion of spectral,thermal and canopy height parameters for improved yield prediction of drought stressed spring barley.Eur. J. Agron. 2016, 78, 44–59. [CrossRef]

23. Thorp, K.R.; Gore, M.A.; Andrade Sanchez, P.; Carmo Silva, A.E.; Welch, S.M.; White, J.W.;French, A.N. Proximal hyperspectral sensing and data analysis approaches for field-based plant phenomics.Comput. Electron. Agric. 2015, 118, 225–236. [CrossRef]

24. Winterhalter, L.; Mistele, B.; Jampatong, S.; Schmidhalter, U. High throughput phenotyping of canopy watermass and canopy temperature in well-watered and drought stressed tropical maize hybrids in the vegetativestage. Eur. J. Agron. 2011, 35, 22–32. [CrossRef]

25. White, J.W.; Andrade Sanchez, P.; Gore, M.A.; Bronson, K.F.; Coffelt, T.A.; Conley, M.M.; Feldmann, K.A.;French, A.N.; Heun, J.T.; Hunsaker, D.J.; et al. Field-based phenomics for plant genetics research. Field CropsRes. 2012, 133, 101–112. [CrossRef]

26. Gee, G.W.; Bauder, J.W. Partical-size analysis. In Methods of Soil Analysis, Part 1. Physical and MineralogicalMethods; ASA-SSSA: Madison, WI, USA, 1986; Chapter 15, pp. 383–411.

27. Zhang, Y.; Schaap, M.G. Weighted recalibration of the Rosetta pedotransfer model with improved estimatesof hydraulic parameter distributions and summary statistics (Rosetta3). J. Hydrol. 2017, 547, 39–53. [CrossRef]

28. Thorp, K.R.; Hunsaker, D.J.; Bronson, K.F.; Andrade-Sanchez, P.; Barnes, E.M. Cotton irrigation schedulingusing a crop growth model and FAO-56 methods: Field and simulation studies. Trans. ASABE 2017,60, 2023–2039. [CrossRef]

29. Hunsaker, D.J.; Barnes, E.M.; Clarke, T.R.; Fitzgerald, G.J.; Pinter, P.J. Cotton irrigation scheduling usingremotely sensed and FAO-56 basal crop coefficients. Trans. ASAE 2005, 48, 1395–1407. [CrossRef]

30. Kalman, R.E. A new approach to linear filtering and prediction problems. J. Basic Eng. 1960, 82, 35–45.[CrossRef]

Remote Sens. 2018, 10, 1682 19 of 19

31. Walter, I.A.; Allen, R.G.; Elliott, R.; Itenfisu, D.; Brown, P.; Jensen, M.E.; Mecham, B.; Howell, T.A.; Snyder, R.;Eching S.; et al. The ASCE Standardized Reference Evapotranspiration Equation; ASCE-EWRI, Reston, VA, USA, 2005.

32. Kullberg, E.G.; DeJonge, K.C.; Chávez, J.L. Evaluation of thermal remote sensing indices to estimate cropevapotranspiration coefficients. Agric. Water Manag. 2017, 179, 64–73. [CrossRef]

c© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open accessarticle distributed under the terms and conditions of the Creative Commons Attribution(CC BY) license (http://creativecommons.org/licenses/by/4.0/).